freeze cannot be finished

[Миша Тюрин <tmihail@xxxxx>]

Hello!
We are experiencing suspicious and very painful case in our top-business-critical database. We have only 7 weeks before emergency stop of the cluster cause wraparound task in autovacuum process can not be finished again and again.

PostgreSQL 9.2.4 on x86_64-unknown-linux-gnu, compiled by gcc-4.4.real (Debian 4.4.5-8) 4.4.5, 64-bit
writing transactions rate ~ 100-200/s (100 000000 per week)

autovacuum_freeze_max_age  1000 000000
vacuum_freeze_min_age             300 000000
vacuum_freeze_table_age           900 000000

vacuum_cost_delay 40 (autovacuum_vacuum_cost_delay -1)
vacuum_cost_limit 700
vacuum_cost_page_dirty 1
vacuum_cost_page_hit 1
vacuum_cost_page_miss 10

Freeze autovacuum ("autovacuum: VACUUM ANALYZE public.items (to prevent wraparound)") runs 4 days 
! but it does not update pg_class.relfrozenxid 

:((

And freeze autovacuum are starting again and again (we have seen it at least 5 times already).
There are no long-running transactions in our cluster.
And we have done manual vacuum freeze on table "items". with vacuum_cost_delay 10 it took 1day and unfortunately it did nothing with pg_class.relfrozenxid .
And we seen (strace into postrges doing vacuum freeze) that vaccuum reads and writes alot into our table's files.

It looks like vacuum at the end cannot update pg_class (maybe).
How should we finish that vacuum freeze?!
Thanks in advance!